[ https://issues.apache.org/jira/browse/AURORA-1868?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15803109#comment-15803109 ]
Mehrdad Nurolahzade commented on AURORA-1868: --------------------------------------------- https://reviews.apache.org/r/55243/ > Evaluate multiple preemption proposals per round > ------------------------------------------------ > > Key: AURORA-1868 > URL: https://issues.apache.org/jira/browse/AURORA-1868 > Project: Aurora > Issue Type: Story > Components: Scheduler > Reporter: Mehrdad Nurolahzade > Assignee: Mehrdad Nurolahzade > Priority: Minor > > {{TaskScheduler}} makes an attempt to preempt already identified candidates > through {{Preemptor}} when it fails to schedule a task. However, > {{Preemptor}} currently evaluates only one candidate per invocation. A > candidate may get vetoed at this point by scheduling filters. If a candidate > fails validation the task group might get penalized by {{TaskGroups}} to give > {{PendingTaskProcessor}} some time to find new preemption candidates; despite > the fact that such a candidate may already exist in slotCache. This penalty > can result in expiration of {{slotCache}} entries, hence slowing down the > overall preemption process. > Consider modifying {{Preemptor}} so that it evaluates all existing preemption > candidates before giving up. This improvement comes at the price of reduced > scheduling throughput. However, in a near capacity cluster, reduced > scheduling throughput is far more acceptable than failure to schedule > critical production services. -- This message was sent by Atlassian JIRA (v6.3.4#6332)